Postscriptum on LLMs: pelicans on bicycles with a twist (danilo.segan.org)

In a recent exploration of the capabilities of large language models (LLMs), an experiment was conducted to generate a PostScript image of a pelican riding a bicycle—an unconventional twist on a popular AI prompt. This investigation highlights how LLMs, especially in coding contexts, are beginning to take on decision-making roles typically handled by human users. By testing the models with a less common vector graphic format compared to the widely used SVG, the author sought to evaluate the models' judgment and adaptability in a scenario with different semantic interpretations. The results were mixed; for example, Gemini 3.7 initially struggled, requiring a correction from an error message to successfully generate the image, while Opus 5 improved through four iterations before successfully rendering a colorized output. This exercise underscores the challenges LLMs face with formats that are less familiar due to limited training data while stressing the importance of human oversight in guiding these models. As LLMs become more integrated into creative workflows, their growing ability to autonomously navigate complex tasks could significantly enhance productivity, although reliability remains a critical aspect to consider as these models evolve.